🧭 Exploratory Data Analysis (EDA): Flights Dataset¶

This notebook documents the Exploratory Data Analysis (EDA) for the Flights dataset, a cornerstone of our Flights MLOps Project. Here, we will dissect the data to uncover critical insights, identify quality issues, and lay a robust foundation for building our flight price prediction model.

🎯 Business Objective¶

To develop and integrate a machine learning-powered flight price prediction tool that provides travelers with real-time, data-driven estimates for future airfare on specific routes. This initiative aims to enhance user engagement and trust by empowering them to make informed booking decisions, identify optimal travel times, and ultimately secure the best possible value. By delivering accurate and transparent fare forecasts, we will position our platform as an indispensable tool for savvy travelers, driving user loyalty and increasing booking conversions.

🔬 Scope of this Analysis¶

The primary goal of this EDA is to deeply understand the dataset to inform the development of a high-performing Regression Model for predicting flight prices.

It's important to note that our data processing decisions are strictly optimized for this modeling objective. For instance, handling duplicate records is tailored to build an unbiased model. Consequently, the resulting dataset may not be suitable for direct business intelligence questions (e.g., calculating total revenue), where every transaction is vital. Our priority is predictive accuracy, not transactional accounting.

📦 Dataset At a Glance: Flights¶

The dataset contains structured records of flight bookings, encompassing travel routes, flight characteristics, pricing, and booking agencies.

🎯 Key Objectives¶

This analysis serves several critical functions in our model development lifecycle:

• Understand Data Structure: Thoroughly examine the schema, data types, and the statistical distributions of all features.
• Identify Data Quality Issues: Pinpoint and quantify problems like missing values, outliers, anomalies, and duplicate records.
• Evaluate Feature Relationships: Analyze how each feature correlates with the target variable (price) and with other features to identify potential predictors.
• Inform Feature Engineering: Use the insights gained to guide data cleaning strategies, transformations, and the creation of new, impactful features for the model.

🔍 Analytical Approach¶

Our exploration will involve a comprehensive analysis using descriptive statistics, data visualizations (histograms, scatter plots, box plots), and correlation studies. The findings are crucial as they will directly inform our feature selection, data transformation techniques, and the overall design of the model training pipeline.

Given the manageable number of features, we will employ a thorough visualization strategy. This includes generating pair plots for numerical features, frequency plots for categorical features, and comparative plots (like box plots) to understand inter-feature relationships and patterns comprehensively.

Let's begin!

Executive Summary: EDA Findings & Modeling Blueprint 🚀¶

Our Exploratory Data Analysis of the Flights dataset reveals a clean, high-quality foundation for modeling after significant deduplication (45% of records removed). The data shows no missing values, and memory usage was halved through optimization.

The strongest predictors for flight price are flightType, Route (from/to) and agency. Furthermore, strong temporal patterns exist, with distinct price and volume seasonality by month and day of the week. The target variable, price, is moderately right-skewed (skewness ~0.53), and distance and time are almost perfectly correlated.

Core Recommendation: Develop a tree-based regression model (e.g., Gradient Boosting) using a strict time-based data split. Key preprocessing steps may include transformation like log-transforming the price, using only one of distance or time to avoid multicollinearity, and engineering features to capture temporal dynamics (e.g., cyclical encoding for month).

1. 🗂️ Data Quality & Preparation¶

• Integrity: The initial dataset of 271,888 records was reduced to 149,484 unique flight instances after removing duplicates based on core flight attributes.This is due to the fact that multiple people on the same date may have booked the same flight on same date. This is crucial to prevent the model from being biased toward common routes. The deduplicated dataset has zero missing values.
• Feature Set: The model will use 8 core features: from, to, flightType, price, time, distance, agency, and date.
• Memory Optimization: Strategic downcasting of data types reduced memory usage from 10.3 MB to 5.1 MB, ensuring better performance in production.

2. 🎯 Key Feature Insights¶

• Target Variable (price):

  • Distribution: Moderately right-skewed (mean $949, median $898).
  • Action: Apply a log transformation (np.log1p) or Other transformations like box-cox, yeo-johnson to normalize the distribution and stabilize variance, which is critical for model performance.

• Primary Predictive Features:

  • flightType: The strongest price differentiator. Mean prices follow a clear hierarchy: firstClass > premium > economic.
  • agency: Significant price variation. 'FlyingDrops' has the highest average price but lowest market share, suggesting a niche, high-margin strategy.
  • Route (from/to): Flights involving Salvador (BH) and Florianopolis (SC) are consistently the most expensive routes.

• Temporal Trends (date):

  • Seasonality: Prices peak during summer months (June-August) and year-end (December).
  • Weekly Pattern: Prices tend to be higher on Fridays and Sundays. Flight volume is highest on Thursdays.
  • Action: Engineer features like month, day_of_week, and is_weekend. Use cyclical encoding for month and day to help the model understand their cyclical nature.

• Numerical Features (distance, time):

  • Multicollinearity: These two features are almost perfectly correlated (~1.0).
  • Action: Use only one of these features (e.g., distance) in the model to avoid multicollinearity.

3. 🤖 Modeling Pipeline Recommendations¶

Import Necessary Libraries and Packages¶

Load The Dataset¶

• Dataset loaded to flights variable as a pandas dataframe

Checking Duplicates¶

check for erroneous duplicates¶

• we need check duplicates with same userCode, date and other features which implies that a user travel entry on same date should not be present multiple times.

Duplicate Values Insights¶

• No Duplicate values were found if we use travelCode and userCode as unique idetifiers.
• No erroneous duplicates found were same userCode would have a duplicate entry on the same date.

Relevance of travelCode and userCode in EDA¶

The travelCode and userCode columns primarily serve as identifiers. They are important for data analysis to gather business insights accurately, separating similar records for different user for same date.

• travelCode: This links related flight segments, such as an outbound and its corresponding return trip. For example:

  | travelCode | userCode | from | to | date | |------------|----------|------------------|--------------------|------------| | 0 | 0 | Recife (PE) | Florianopolis (SC) | 09/26/2019 | | 0 | 0 | Florianopolis (SC)| Recife (PE) | 09/30/2019 |

• userCode: This links flights to specific users, enabling connections to the users.csv data.

• It's possible for different travelCode or userCode entries to correspond to flights with identical details (route, price, time, distance, agency,date). This can happen if multiple users book the same flight. Including these identifiers directly in analyses focused on flight attributes without aggregation could potentially cause Bias in a model by over-representing frequently booked flight configurations.

Checking And Handling duplicate values after removing unique identifiers(travelCode and userCode)¶

• This suggest many of the duplicates were hidded due to unique identifiers

✈️ Addressing Duplicate Flight Records¶

Our analysis revealed a significant number of duplicate entries within the dataset after removal of unique identifiers:

• Substantial Duplicate Presence: Approximately 45.02% of the flight records were identified as duplicates. This high percentage is attributed to multiple unique travelCode and userCode entries corresponding to identical flight details (route, timing, price, etc.), likely representing different bookings for the same underlying flight service on same date by different users.

• Critical Preprocessing Step: Duplicate Removal:

  • Although for data analysis to accurately answer business questions like total number of sales etc its very important to not remove these duplicate entries as they are from a different user however for model building and doing eda specifically for eda we have to remove these duplicates carrying same information.
  • Rows with different date even if rest of the features are same will be preserved as they shows time trend with price. 🚮 It is crucial to remove these duplicate records before proceeding with further analysis or model training.
  • Impact on Analysis: Retaining duplicates would skew descriptive statistics and visualizations, leading to inaccurate interpretations of flight patterns and pricing.
  • Impact on Modeling: For machine learning, these duplicates can introduce significant bias, leading to an overestimation of model performance on seen data and poor generalization to new, unseen flight queries. Removing them ensures model integrity and more reliable predictions.

📑 Automated Basic Stats Report with ydata-profiling¶

• This will calculate basic stats about flights dataset and wont perform any extensive eda which may cause data leakage and bias.

Column Information, Feature creation and DataType Opimization¶

Breaking Date Column¶

• Extracting Year, Month , Day and Day name from Date column

Optimize Dataypes¶

Column Information and Data Type Optimization¶

Effective data management is crucial for efficient analysis and model training. This section details the steps taken to inspect the dataset's structure, understand its column types, and optimize these types for better memory usage and computational performance.

Initial Data Structure¶

Upon loading and initial cleaning (including duplicate removal), the dataset comprised 149,484 flight records. A preliminary check using flights.info() revealed the following data types:

• Object Types: from, to, flightType, agency, date. These are typically strings and require careful handling.
• Float64 Types: price, time, distance. These are standard 64-bit floating-point numbers.

The flights.nunique() command was used to assess the number of unique values in each column, providing an early indication of which columns might be suitable for conversion to more memory-efficient types like category.

Date Feature Engineering and Standardization¶

The date column, initially an object type, is vital for time-series analysis and understanding temporal patterns. The following transformations were applied:

• Conversion to Datetime: The date column was converted to Pandas' datetime64[ns] format using pd.to_datetime(flights["date"], errors="coerce"). This standardization is essential for performing date-specific operations.
• Feature Extraction: To leverage temporal information more directly, new columns were extracted from the date column:
  • month: The month of the flight (e.g., 1 for January, 12 for December).
  • year: The year of the flight.
  • day: The day of the month.
  • day_name: The name of the day of the week (e.g., Monday, Tuesday).

These new features allow for more granular analysis of trends related to specific times of the year or days of the week.

Data Type Optimization Strategy¶

To enhance performance and reduce the memory footprint of the dataset, the following data type optimizations were implemented using a custom optimize_dtypes function and specific column conversions:

• Categorical Conversion for Low Cardinality Features:

  • Columns such as from, to, flightType, agency, and the newly created year were identified as having a relatively small number of unique values (low cardinality).
  • These were converted from object or numerical types to the category dtype. This is highly beneficial as Pandas stores category data more efficiently by using integer codes internally, significantly reducing memory usage for repetitive string data.

• Downcasting Numerical Features:

  • The numerical columns price, time, and distance, initially float64, were downcast to float32. This conversion reduces the memory required for each floating-point number by half, often without a practical loss of precision for the type of data represented (e.g., flight prices, durations, distances).

• Optimizing Extracted Date Components:

  • The newly created month and day columns, representing integer values within a small range, were optimized to uint8 (unsigned 8-bit integer). However for Modeling month and day can be converted to cyclic.
  • The year column, before its conversion to category, was suitable for uint16 (unsigned 16-bit integer).

Impact of Optimization¶

The combined effect of these data type optimizations was substantial:

• Memory Reduction: The dataset's memory usage was significantly reduced. Initially, after basic loading and cleaning, the memory footprint was over 10.3 MB. After the optimization process, this was brought down to approximately 5.1 MB.
• Performance Gains: Smaller data types lead to faster processing for many Pandas operations and can accelerate the training of machine learning models.

By thoughtfully optimizing data types, we ensure that the dataset is not only memory-efficient but also primed for faster and more effective subsequent analytical and modeling tasks. This step is a best practice in any data science workflow, particularly when dealing with larger datasets.

5 Number Summary¶

Descriptive Statistics: The Five-Number Summary and Categorical Overview¶

To gain a foundational understanding of the dataset's characteristics, descriptive statistics were computed for both numerical and categorical features. This includes the five-number summary (minimum, first quartile (Q1), median (Q2), third quartile (Q3), and maximum) for numerical variables, and frequency distributions for categorical variables.

Numerical Feature Summary¶

The flights.describe(include="number") method provided key statistical measures for the numerical columns: price, time, distance, month, and day.

• price (Target Variable):

  • Mean: Approximately 949.42 currency units.
  • Standard Deviation: Approximately 358.45, indicating a moderate spread in prices.
  • Minimum: 301.51 currency units.
  • 25th Percentile (Q1): 669.83 currency units.
  • Median (50th Percentile): 898.04 currency units. The median being slightly lower than the mean suggests a mild positive (right) skew in the price distribution, implying a larger number of flights at lower price points with a tail of higher-priced flights.
  • 75th Percentile (Q3): 1214.82 currency units. The interquartile range (IQR = Q3 - Q1) is approximately 544.99.
  • Maximum: 1754.17 currency units.
  • Interpretation: Flight prices exhibit a considerable range, with most falling within the IQR. The skewness indicates that while the bulk of prices are below 1000 units, there are instances of significantly higher prices.

• time (Flight Duration in Hours):

  • Mean: Approximately 1.36 hours.
  • Standard Deviation: Approximately 0.55 hours, reflecting moderate variability in flight durations.
  • Minimum: 0.44 hours (approximately 26 minutes).
  • Median: 1.44 hours. The median is very close to the mean, suggesting a relatively symmetric distribution of flight times.
  • Maximum: 2.44 hours.
  • Interpretation: Flight durations are generally clustered around the mean, with a limited range from short hops to medium-length flights.

• distance (Flight Distance):

  • Mean: Approximately 524.68 units (e.g., kilometers).
  • Standard Deviation: Approximately 211.00 units, indicating a fairly wide spread in flight distances.
  • Minimum: 168.22 units.
  • Median: 555.74 units. The median being slightly higher than the mean suggests a slight negative (left) skew, though the distribution is close to symmetrical.
  • Maximum: 937.77 units.
  • Interpretation: The dataset covers a diverse range of flight distances, from short regional flights to longer domestic routes.

• month and day:

  • These features, derived from the date, represent the month of the year (1-12) and the day of the month (1-31). Their distributions are as expected, covering the full range of possible values. The mean for month is around 6.58, and for day is around 15.76, indicating data is spread across the calendar.

Categorical Feature Summary¶

The flights.describe(include=["object", "category"]) method provided insights into categorical columns:

• from (Origin City/Airport):

  • Unique Values: 9 distinct origin locations.
  • Top (Most Frequent): "Florianopolis (SC)" with 18,800 occurrences.
  • Interpretation: Indicates a limited set of departure points, with Florianopolis (SC) being a primary hub.

• to (Destination City/Airport):

  • Unique Values: 9 distinct destination locations.
  • Top (Most Frequent): "Florianopolis (SC)" with 31,438 occurrences.
  • Interpretation: Similar to origins, destinations are concentrated, with Florianopolis (SC) also being a major arrival hub. The higher frequency as a destination suggests it might be a more common endpoint or a central connecting point.

• flightType (Service Class):

  • Unique Values: 3 distinct flight types.
  • Top (Most Frequent): "firstClass" with 64,169 occurrences.
  • Interpretation: First class is the most represented service type in this dataset, followed by other classes like economic and premium.

• agency (Booking Agency):

  • Unique Values: 3 distinct agencies.
  • Top (Most Frequent): "Rainbow" with 64,167 occurrences.
  • Interpretation: The "Rainbow" agency handles the largest volume of bookings in this dataset.

• year:

  • Unique Values: 5 distinct years.
  • Top (Most Frequent): 2020 with 55,095 occurrences.
  • Interpretation: The data spans five years, with 2020 having the highest representation of flight records.

• day_name (Day of the Week):

  • Unique Values: 5 distinct day names (e.g., Monday to Friday, as weekends were noted to be excluded).
  • Top (Most Frequent): "Thursday" with 52,115 occurrences.
  • Interpretation: Flights are distributed across the weekdays, with Thursday showing the highest frequency in this particular dataset.

Overall Summary¶

The descriptive statistics reveal a dataset characterized by flights primarily within a domestic or regional scope, given the distance and time ranges. Prices show variability and a slight positive skew. Key hubs, flight classes, and agencies are evident from the categorical data. The temporal distribution indicates that the data is spread across several years, with specific years and days of the week showing higher flight volumes. These initial statistics are crucial for understanding data distributions, identifying potential outliers, and informing subsequent feature engineering and modeling decisions.

Date Analysis for Splitting Criteria¶

• To prevent data leakage and ensure robust model performance, it is essential to use only historical (past) data during exploratory data analysis (EDA) and model training phases. Future data should be strictly reserved for validation and testing.
• Implementing date-based splits simulates real-world scenarios by maintaining temporal order, enabling realistic and unbiased model evaluation while preserving the integrity and generalizability of your results.

✈️ Flights Data Split Strategy¶

Note!: Holdout set was not created in production project instead train:70%, validation:15%, test:15% configuration was used.

📝 Initial Split Insights¶

• Chronological splits are fundamental to prevent data leakage and accurately reflect real-world model deployment scenarios where predictions are made on future, unseen data.
• Seasonality is preserved: Each split is designed to cover multiple months, aiming to capture various holiday and travel patterns.
• Balanced validation and test sets (approx. 15% each) ensure reliable model tuning and an unbiased final evaluation.
• Production simulation (🚦) utilizes the most recent data, offering a realistic preview of model performance during retraining cycles and aiding in drift monitoring.

💡 Why This Chronological Split is Crucial (Evidence-Based Justification)¶

A date-based (chronological) split is not just a best practice; it's essential for this dataset due to observable temporal patterns:

• Mimics Real-World Deployment 🌎: The core principle is to train the model on past data and evaluate its ability to predict future outcomes. This directly mirrors how the model would be used in a production environment.
• Captures Evolving Trends & Seasonality 📈:
  • The "Flights per Month" bar chart clearly shows significant seasonality in flight volume. For instance, October, December, and January exhibit peak flight numbers, while months like June, July, August, and September show relative dips. A random split could create imbalanced representations of these crucial periods across train, validation, and test sets, leading to a model that performs poorly on unseen seasonal data.
  • The average price table (month/year) reveals that prices are not static. There are noticeable year-over-year price trends (e.g., prices in 2022 and 2023 are often higher than in 2020/2021 for the same month) and distinct monthly price fluctuations within each year (e.g., July 2023 shows a significantly higher average price). A chronological split ensures the model learns these temporal price dynamics and how seasonality might interact with longer-term price evolution.
• Prevents Lookahead Bias (Data Leakage) 🕵️: Randomly shuffling data that has a temporal component would allow the model to inadvertently learn from future information during training, leading to overly optimistic performance metrics that don't hold up in real-world scenarios.
• Ensures Robust Model Evaluation & Lifecycle Management 🔄: By testing on a distinct future period, we get a more reliable estimate of the model's generalization capabilities. The "Production Sim" split further allows for testing how the model adapts to the very latest data patterns, which is vital for ongoing monitoring and retraining decisions.

Missing Values¶

🕵️‍♂️ Missing Value Assessment¶

• ✅ No Missing Data Found: A thorough check for missing values across all columns and rows confirmed that the current dataset is complete. There are no missing values in any cell.

• Implication for Current Analysis: This is a positive finding for the initial EDA and model prototyping phases, as it simplifies preprocessing by eliminating the need for immediate imputation strategies for the existing data.

• Consideration for Production Pipelines: While the current dataset is clean, it's crucial to design production MLOps pipelines to be robust against potential missing data in future incoming flight records. Therefore, strategies for handling missing values (e.g., imputation or flagging) should still be incorporated into the production data preprocessing steps.

¶

Analysis on Training Set¶

Spitting Data into Training, Valdiation, Testing and Holdout(For Production Simulation) Sets¶

Note!: Holdout set was not created in production project instead train:70%, validation:15%, test:15% configuration was used.

📑 Extensive Automated EDA for Training Set¶

• This will generate a extensive EDA report fot the training set.

Insights from the Report¶

• Distance shows a strong positive correlation with both price and flight time.
• Month and year are highly correlated, reflecting the chronological structure of the data.
• Price is strongly correlated with both distance and flight time, indicating that longer and longer-duration flights tend to be more expensive.
• Flight time is also highly correlated with both distance and price, suggesting that longer flights cover greater distances and cost more.
• Year and month exhibit a strong correlation, as expected in time series data.

Detailed report can be found in reports/eda folder which can be viewed as html file in browser.

5 Number Summary¶

✈️ Five Number Summary Insights¶

• Flight Time ⏱️:

  • Ranges from 0.44 to 2.44 hours, with an average of 1.35 hours.
  • The median is 1.44 hours, and most flights are between 0.85 and 1.76 hours, indicating a focus on short to medium-haul routes.

• Distance 📏:

  • Flights cover 168 km to 938 km, with an average of 520 km.
  • The median distance is 556 km, and most flights fall between 328 km and 677 km, showing a concentration on regional travel.

• Month & Day 📅:

  • Flights are distributed across all months (1–12) and days (1–31), with the median flight in June on the 16th day.
  • The average month is about June (6.4), and the average day is the 16th, suggesting relatively even distribution throughout the year and month.

• Price 💸:

  • Prices range from 301.51 to 1,754.17, with an average of 944.79.
  • The median price is 889.07, and most flights are priced between 667.00 and 1,209.04.
  • The right-skewed distribution indicates a few high-priced flights.

• Categorical Features 🏷️:

  • 9 unique origins and destinations, with Campo Grande (MS) as the most common origin and Florianopolis (SC) as the top destination.
  • FirstClass is the most frequent flight type, and Rainbow is the leading agency.
  • The year 2020 dominates the data, with 55,095 flights.

• Date Range 📆:

  • Data spans from 2019-09-26 to 2021-05-22.
  • The median flight date is 2020-07-02, and the average date falls in mid-2020, indicating a concentration of flights during this period.

Overall:
The dataset primarily features short to medium-haul flights, with a strong presence of first-class tickets and a concentration of flights in 2020. Prices and distances show moderate variability, and the data is well-distributed across months and days, supporting robust time-based analysis.

Skew in Each Numerical Column¶

Skewness in Numerical Features¶

Skewness measures the asymmetry of a feature's distribution. Understanding this is crucial as it can guide feature transformation choices and influence model performance. The following skewness values were observed for the numerical columns:

• price (Target Variable) 💰:

  • Skewness Value: +0.353881
  • Interpretation: Exhibits a mild positive (right) skew. This indicates that while many flight prices are concentrated at the lower end of the distribution, there's a tail extending towards higher prices. This suggests more flights are relatively inexpensive, with fewer, but notably more expensive, options.

• time (Flight Duration) ⏱️:

  • Skewness Value: -0.055489
  • Interpretation: The distribution of flight duration is very close to symmetrical. The skewness value is slightly negative but very near zero, implying that flight times are quite evenly distributed around the average, without a significant bias towards shorter or longer durations.

• distance (Flight Distance) 🗺️:

  • Skewness Value: -0.059037
  • Interpretation: Similar to flight time, the distribution of flight distance is very close to symmetrical. The slightly negative skewness value is minimal, indicating a balanced distribution of flight distances around their mean.

• month (Month of Flight) 🗓️:

  • Skewness Value: +0.065595
  • Interpretation: The distribution of flights across months is very close to symmetrical. The small positive skewness value suggests a fairly even spread of flight records throughout the months of the year, without a strong lean towards earlier or later months in the overall dataset distribution (though seasonal peaks exist, skewness here refers to the shape of the count distribution itself).

• day (Day of Month) 🔢:

  • Skewness Value: +0.007979
  • Interpretation: The distribution of flights by the day of the month is highly symmetrical. The skewness value is extremely close to zero, indicating that flight occurrences are very evenly spread across the days of the month.

Summary of Skewness Insights:

The target variable, price, displays a mild positive skew, a common characteristic for pricing data. The primary predictor candidates, time and distance, along with the temporal features month and day, all exhibit distributions that are very close to symmetrical. This general lack of strong skew in most predictors is favorable, as many machine learning algorithms perform optimally with or assume near-symmetrical data distributions. For the price variable, its mild skew might still warrant consideration for transformation (e.g., log transformation) depending on the chosen modeling approach to potentially improve model performance and linearity assumptions.

Univariate Analysis¶

📊 Univariate Analysis of Numerical Features¶

Distribution Of Numerical Columns¶

📊 Distribution Analysis Main Insights¶

💰 Price¶

The price distribution is multimodal and exhibits a slight right skew. The presence of several peaks clearly indicates multiple common price ranges. These variations can be attributed to factors such as different flight classes, varying routes, or periods of high demand. Prices are distributed from approximately 300 to 1800, without extreme outliers.

🗓️ Day¶

The distribution of flights across days 1–31 is nearly uniform. This uniformity demonstrates no particular day-of-the-month bias in the flight data.

📅 Month¶

Flight frequencies vary significantly by month. Months such as January (❄️), February, March, April, October (🍂), November, and December show higher flight frequencies. Conversely, June, July (☀️), August, and September have lower frequencies. This pattern strongly indicates seasonality in flight operations, with more flights occurring in the winter and spring months and fewer during the summer and early fall.

✈️ Distance¶

The distribution of flight distances is multimodal, characterized by several distinct peaks. This pattern shows that flights are predominantly clustered around certain common distances, which likely correspond to popular or standard flight routes.

⏱️ Time¶

The distribution of flight durations is also multimodal. A significant peak is observed around 1.5 (presumably hours), indicating that flight durations of this length are considerably more common. This reflects typical times for standard routes.

📝 Summary of Univariate Insights¶

The data reveals clear and actionable patterns:

• Seasonality Trends 📈: Flight activity varies significantly by month.
• Route Clustering 🗺️: Flight distances and durations show distinct clustering, indicative of popular routes.
• Pricing Bands 🏷️: Multiple distinct price ranges are evident in the data.

No major data quality issues are apparent from this initial univariate examination.

Outlier Detection Using Boxplots¶

📦 Box Plot Analysis Insights¶

💰 Price¶

• Median: ~889
• IQR: 542 (Q1: 667, Q3: 1209)
• Range: 301 (min) to 1754 (max) Insight: The price distribution is fairly wide, underscored by a large interquartile range. While no extreme outliers are present, this spread signifies considerable price variability. This variation likely stems from diverse factors such as different routes, flight classes, or booking times.

🗓️ Day¶

• Median: 16
• IQR: 15 (Q1: 8, Q3: 23)
• Range: 1 to 31 Insight: Days are uniformly distributed, as expected for calendar days. The absence of outliers indicates a balanced sampling across the month.

📅 Month¶

• Median: 6
• IQR: 7 (Q1: 3, Q3: 10)
• Range: 1 to 12 Insight: Months are also well-distributed, with no outliers detected. This suggests the dataset comprehensively covers the full year.

✈️ Distance¶

• Median: ~556
• IQR: ~349 (Q1: 328, Q3: 677)
• Range: 168 to 938 Insight: Flight distances vary widely, although the majority fall between 328 and 677 units (likely kilometers or miles). No extreme outliers are observed, but the spread indicates a mix of short and medium-haul flights.

⏱️ Time¶

• Median: 1.44
• IQR: 0.91 (Q1: 0.85, Q3: 1.76)
• Range: 0.44 to 2.44 Insight: Flight durations are mostly concentrated between 0.85 and 1.76 (likely hours). The distribution is relatively tight with no outliers, suggesting that most flights are of similar duration.

📊 Overall Box Plot Summary¶

• The data demonstrates a good distribution across time-related features (Day, Month) and categories.
• Price and Distance exhibit the most variability, highlighting them as potentially significant factors for predictive modeling.
• The absence of significant outliers across these features points towards good data quality. ✅
• The observed spread in price and distance may reflect different flight types or routes, which could serve as useful features for prediction tasks.

Univariate Categorical Analysis¶

📊 Categorical Variable Bar Plot Insights¶

🗓️ day_name¶

• Thursday shows significantly higher flight counts, almost double compared to other days.
• Monday, Friday, Saturday, and Sunday have similar, much lower counts. Insight: A strong day-of-week effect is evident, with Thursday emerging as a peak travel day. This could be influenced by business travel patterns or specific airline scheduling.

📅 year¶

• 2020 has the highest number of flights, followed by 2021 and 2019.
• 2022 and 2023 have zero records. Insight: The dataset is primarily concentrated in the 2019–2021 period, with a notable spike in 2020. This distribution may reflect the data collection's specific timeframe or be indicative of real-world events (e.g., initial phases or disruptions related to COVID-19).

🏢 agency¶

• CloudFly and Rainbow show similar, high flight counts.
• FlyingDrops has a considerably lower number of flights. Insight: The market appears to be dominated by two main agencies (CloudFly and Rainbow), with FlyingDrops playing a more minor role.

✈️ flightType¶

• firstClass is the most common flight type.
• economic and premium classes follow, with nearly equal, lower counts. Insight: There is a higher proportion of first-class flights in this dataset. This is somewhat unusual and may reflect the specific focus of the dataset or the primary offerings of the included agencies.

🎯 to (Destination)¶

• Florianopolis (SC) is the most frequent destination. 🏖️
• Aracaju (SE), Campo Grande (MS), and Recife (PE) also have high destination counts.
• Rio de Janeiro (RJ) and Salvador (BH) have the lowest counts as destinations. Insight: Certain destinations are significantly more popular than others. This popularity could be driven by tourism, business activities, or these locations serving as major airline hubs.

🛫 from (Origin)¶

• Flight origins are more evenly distributed compared to destinations.
• Aracaju (SE), Brasilia (DF), Campo Grande (MS), and Recife (PE) have similar, relatively high origination counts.
• Rio de Janeiro (RJ) and Salvador (BH) are less frequent as origin points. Insight: Flights originate fairly evenly from most listed cities. However, a comparison with destination data reveals that some cities are more prominent as destinations than as origins.

📜 Overall Categorical Insights Summary¶

• Strong patterns are observed in day-of-week, year, agency, and destination distributions.
• Thursday (for day_name) and the year 2020 (for year) stand out as notable peaks within their respective categories.
• The distributions for agency and flight type suggest a non-uniform market structure and service offering.
• These categorical features, with their distinct patterns, are highly likely to be important predictors in any subsequent modeling tasks. ✨

Bivariate Analysis¶

Pairplot¶

• Ploting scatter plots between all numerical columns to see the relation ship.

Correlation Analysis using Corr Matrix And Heatmap¶

🔗 Correlation Analysis Insights¶

✈️⏱️ Strong Correlation: time and distance¶

• Correlation Coefficient ≈ 1.0 (specifically 0.99999) Insight: Flight time and distance are almost perfectly correlated in this dataset. This indicates an essentially linear relationship, meaning as distance increases, flight time increases proportionally. This is expected, as longer flights naturally cover more ground, assuming relatively consistent cruising speeds.

💰 Moderate Correlation with price¶

• price vs time: 0.65
• price vs distance: 0.65 Insight: Both flight time and distance show a moderate positive correlation with price. This suggests that longer and farther flights generally tend to be more expensive. However, the correlation isn't perfect, implying that other factors (such as flight class, agency, demand, or booking time) also significantly influence the final price.

🗓️ Very Weak or No Correlation with month and day¶

• All correlations involving month and day with price, distance, or time are near zero. Insight: The month and day of the flight do not exhibit a linear relationship with price, distance, or time. Any influence these temporal features have on pricing or flight characteristics is likely non-linear or categorical (e.g., specific effects of holidays, weekends, or seasonal demand peaks not captured by a simple linear trend).

📊 Pairplot Visuals¶

• Scatter Plots:
  • The plot of time vs distance displays a tight, clear linear pattern, reinforcing their strong correlation.
  • Plots for price vs time and price vs distance show discernible positive trends, but with a greater degree of spread (variance), aligning with the moderate correlation values.
  • Other variable pairs generally show no clear linear patterns.
• Distributions:
  • The individual distribution plots for each variable along the diagonal of the pairplot match the shapes observed in the earlier histogram analysis.

📝 Summary of Correlation Insights¶

• Key Numerical Drivers 🔑: Distance and time are identified as the primary numerical features linearly influencing price.
• Multicollinearity Alert ⚠️: time and distance are highly redundant due to their near-perfect correlation. For modeling purposes, it's likely that only one of these features will be necessary to avoid multicollinearity.
• Handling Other Features: Month and day will likely need to be treated as categorical features or used to create interaction terms with other features to effectively capture their impact on the target variable.

These insights are crucial for guiding feature selection and feature engineering efforts in the development of predictive models.

Boxplots For Numerical Insights From Categorical Columns¶

✈️ Analysis of Flight Characteristics by Category¶

This section details how flight time, distance, and price vary across different categorical features.

🗓️ By Day of the Week¶

• Flight Duration & Distance 🕰️📏: Flight times and distances exhibit high consistency across all days of the week, with only minor variations observed.
• Ticket Prices 💰: Prices generally remain stable across the days. A notable dip in prices occurs on Thursdays, which aligns with the observation that Thursdays experience higher flight volumes.

📅 By Year¶

• Flight Duration, Distance & Price Over Time ⏳: Analysis of data from 2019 through 2021 indicates stability in flight times, distances, and prices. Median values and data spreads for these variables show minimal change year-over-year, suggesting that despite potential seasonal trends, the average price point remained consistent during this period.

🏢 By Agency¶

• Flight Duration & Distance 🚀: The distributions for flight time and distance are nearly identical across all analyzed agencies (CloudFly, FlyingDrops, and Rainbow). No single agency demonstrates a significant deviation in terms of offering consistently faster or longer flights.
• Ticket Prices 💸: A clear distinction in pricing is observed: FlyingDrops presents noticeably higher median prices compared to CloudFly and Rainbow. The latter two agencies offer prices that are largely similar to each other.

💺 By Flight Type¶

• Flight Duration & Distance 🗺️: Flights categorized as Economic, firstClass, and premium show almost identical distributions for both flight time and distance. The class of service does not correlate with significant differences in route length or flight duration.
• Ticket Prices 🏷️: As anticipated, firstClass flights are the most expensive. Premium flights are priced in the mid-tier, while economic class flights are the least expensive. This confirms standard airline pricing structures.

🌍 By Origin (from) and Destination (to)¶

• Flight Duration ⏱️:
  • Flights originating from or destined for Salvador (BH) and Florianopolis (SC) generally have longer durations.
  • Conversely, flights involving Natal (RN) and Sao Paulo (SP) tend to have shorter flight times.
• Flight Distance ✈️:
  • Consistent with duration, routes to/from Salvador (BH) and Florianopolis (SC) also cover the longest distances.
  • Routes involving Natal (RN) and Sao Paulo (SP) are typically the shortest in terms of distance.
• Ticket Prices 💲:
  • Journeys involving Salvador (BH), Florianopolis (SC), and Aracaju (SE) are associated with higher ticket prices.
  • Flights to or from Sao Paulo (SP) and Brasilia (DF) are generally among the least expensive.

This revised version should be more suitable for your documentation needs.

Barplots Between Categorical columns and Numerical Agregations(Mean, Mode, Median, Sum)¶

📊 Categorical vs. Numerical Aggregations: Bar Plot Insights¶

This section details insights derived from bar plots showing aggregated numerical values (sum and mean) for different categorical features.

➕ Sum Aggregation Insights¶

These insights reflect the total sums (e.g., total sales, total distance covered) for each category.

🛫 By Origin (from)¶

• Total Price (Sales Volume) 💰:
  • The highest total sales were generated from flights originating in Brasilia (DF), closely followed by Campo Grande (MS) and Aracaju (SE).
  • The lowest total sales were from flights originating in Salvador (BH). This contrasts with its distance/time metrics, potentially indicating fewer, but possibly more expensive, tickets or simply fewer flights overall from this origin in the dataset.
• Total Time & Distance 🕰️📏:
  • These two metrics show similar trends. The greatest total flight distance and time were covered by flights originating from Florianopolis (SC), followed by Aracaju (SE).

🎯 By Destination (to)¶

• Total Price (Sales Volume) 💰:
  • Flights destined for Florianopolis (SC) generated the highest total sales by a significant margin, followed by Aracaju (SE).
  • The lowest total sales were for flights destined to Rio de Janeiro (RJ).
• Total Time & Distance 🕰️📏:
  • The trend for total time and distance precisely mirrors the trend observed for total price, with Florianopolis (SC) leading.

✈️ By Flight Type¶

• Total Time & Distance 🕰️📏:
  • A consistent pattern is observed: firstClass flights account for the largest total flight time and distance, followed by economic class, and then closely by premium class.
• Total Price (Sales Volume) 💰:
  • Total sales are highest for firstClass flights, followed by premium, and then economic class.

🏢 By Agency¶

• Total Time & Distance 🕰️📏:
  • FlyingDrops accounts for the lowest total flight time and distance.
  • Rainbow shows the highest total time and distance by a substantial margin compared to FlyingDrops, and is closely followed by CloudFy.
• Total Price (Sales Volume) 💰:
  • The sales pattern mirrors the time and distance trends: Rainbow generated the highest total revenue, followed by CloudFy, with FlyingDrops earning the least.

📅 By Year¶

• Total Time & Distance 🕰️📏:
  • The year 2020 recorded the highest total flight time and distance by a significant margin, followed by 2021 and then 2019. This lower figure for 2019 may be attributable to the dataset starting in the 9th month of that year.
• Total Price (Sales Volume) 💰:
  • Total sales follow a similar trend to total time and distance, with 2020 being the highest.

🗓️ By Day of the Week (day_name)¶

• Total Time, Distance, & Price 🕰️📏💰:
  • Thursday exhibits the highest total flight distance, total flight time, and total sales. This could be linked to the observation that flights are comparably cheaper on Thursdays, potentially driving higher volume, or other contributing factors.

⚖️ Mean Aggregation Insights¶

These insights reflect the average values (e.g., average price, average flight duration) for each category.

🛫 By Origin (from)¶

• Mean Time & Distance 🕰️📏:
  • These metrics show similar trends. Flights originating from Salvador (BH) have the highest mean flight time and distance, followed by Florianopolis (SC) and Aracaju (SE).
• Mean Price 💰:
  • The highest mean price is for flights originating from Salvador (BH), followed by Brasilia (DF) and Rio de Janeiro (RJ).

🎯 By Destination (to)¶

• Mean Time & Distance 🕰️📏:
  • A similar pattern is observed: Salvador (BH) has the highest mean flight time and distance for arrivals, followed by Florianopolis (SC) and Aracaju (SE).
  • The lowest mean time and distance are for flights arriving at Sao Paulo (SP) and Brasilia (DF).
• Mean Price 💰:
  • The highest mean price is for flights destined to Salvador (BH), followed by Florianopolis (SC) and Aracaju (SE).
  • The lowest mean price is for flights arriving at Sao Paulo (SP).

✈️ By Flight Type¶

• Mean Time & Distance 🕰️📏:
  • Mean flight time and distance are similar across all flight types (Economic, firstClass, premium).
• Mean Price 💰:
  • firstClass flights have the highest mean price, followed by premium, and then economic class.

🏢 By Agency¶

• Mean Time & Distance 🕰️📏:
  • Mean flight time and distance show similar values across the different agencies.
• Mean Price 💰:
  • FlyingDrops has the highest mean price.
  • Rainbow and CloudFy have similar, lower mean prices compared to FlyingDrops.

📅 By Year¶

• Mean Time, Distance, & Price 🕰️📏💰:
  • Mean flight distance, mean price, and mean time are all relatively similar across the years (2019-2022, where data is available).

🗓️ By Day of the Week (day_name)¶

• Mean Time & Distance 🕰️📏:
  • Mean flight time and distance are marginally lowest on Thursday. Values for other days of the week are quite similar to each other.
• Mean Price 💰:
  • The mean price is lower on Thursday compared to other days of the week.

EDA Summary And Model Pipeline Recommendations for Flight Price Prediction¶

A comprehensive summary of the Exploratory Data Analysis (EDA), integrating detailed findings from univariate, bivariate, and multivariate analyses, including aggregations and temporal trend assessments. This summary provides actionable insights crucial for developing a robust and accurate flight price prediction model.

1. 🗂️ Data Understanding & Quality Assessment¶

• Schema & Data Types:

  • The dataset initially includes travelCode and userCode. For modeling purposes, the 8 core features are: from (Origin City), to (Destination City), flightType (Service Class), price (Target Variable), time (Flight Duration), distance (Flight Distance), agency (Airline Agency), and date (Flight Date).
  • Categorical Features: from, to (9 unique cities each), flightType (3 unique values: 'economic', 'premium', 'firstClass'), agency (3 unique values). These, along with extracted year and day_name, were converted to category dtype.
  • Numerical Features: price, time, distance were downcast to float32. Extracted month and day were optimized to uint8.
  • Temporal Feature: date was converted to datetime64[ns] objects, and features like year, month, day, day_name were extracted and month, day can be converted to cyclic encoding for better performance.

• Dataset Size & Integrity:

  • Initial State: The raw dataset contained 271,888 rows.
  • Duplicate Records Analysis:
    • No duplicate values were found when using travelCode and userCode as unique identifiers, confirming the integrity of each individual record.
    • Furthermore, no erroneous duplicates were found where the same userCode had a duplicate entry for the same flight leg (i.e., on the same date and time).
    • To prepare the data for modeling, the identifiers were removed to assess duplication based on core flight attributes. This revealed that 45.02% of the records were duplicates.
    • Critical Justification: Although for data analysis to accurately answer business questions like total number of sales etc its very important to not remove these duplicate entries as they are from a different user however for model building and doing eda specifically for eda we have to remove these duplicates carrying same information. This prevents the model from being biased towards more frequently traveled flight routes.
    • Rows with different date even if rest of the features are same will be preserved as they shows time trend with price.
  • Deduplicated Dataset: After removal of these duplicates, the dataset used for EDA and modeling consists of 149,484 unique flight instances.

• Missing Values Assessment:

  • ✅ The current dataset (after deduplication and processing) shows no missing values in any column, as confirmed by flights.info() showing non-null counts equal to the total number of entries for all columns.
  • Proactive Strategy: Despite current completeness, production pipelines must include robust imputation strategies (e.g., mean/median/mode imputation for numerical, constant/mode for categorical) to handle potential missing data in new, incoming flight records.

• Memory Efficiency:

  • Strategic downcasting of numerical types (e.g., float64 to float32, appropriate uint types for date components) and conversion of object strings to category dtype resulted in a substantial memory usage reduction. The memory footprint for the 149,484 records was reduced from over 10.3 MB (with initial dtypes after loading and dropping original ID columns) to approximately 5.1 MB after all optimizations.
  • Production Implication: Utilizing optimized data types is crucial for faster processing, reduced memory footprint, and improved scalability in production environments.

2. 🎯 Target Variable (price) Analysis¶

• Distributional Characteristics (from flights.describe() and skewness calculation):

  • Central Tendency: Mean: $949.42, Median: $898.04. The mean being higher than the median indicates a positive skew.
  • Spread & Range: Min: $301.51, Max: $1754.17. Standard Deviation: $358.45.
  • Quartiles: 25th Percentile (Q1): $669.83, 75th Percentile (Q3): $1214.82. Interquartile Range (IQR): $544.99.
  • Skewness: The price distribution is slightly right-skewed (skewness ~0.53), with a longer tail towards higher prices. This was confirmed by visual inspection (histograms, box plots in the notebook) and skewness statistics.

• Implications for Modeling:

  • The presence of higher-priced outliers and the right skew can disproportionately influence models sensitive to data distribution (e.g., linear regression).
  • Action: Consider applying a log transformation (e.g., np.log1p) to the price variable to normalize its distribution, stabilize variance, and potentially improve model performance and interpretability.

3. 🏗️ Feature Analysis, Relationships & Engineering Insights¶

• Numerical Features (distance, time):

  • Correlation with price: Both distance and time exhibit a moderately strong positive correlation with price (correlation coefficients around 0.65, as seen in the correlation heatmap). Longer flights (in distance or duration) generally cost more.
  • Inter-Correlation (Multicollinearity): distance and time are almost perfectly positively correlated with each other (correlation coefficient ~1.0 from the heatmap). They essentially convey redundant information.
    • distance: Mean: 524.68 units, Median: 555.74 units. Range: 168.22 to 937.77 units. Skewness ~0.00 (nearly symmetric).
    • time: Mean: 1.36 hours, Median: 1.44 hours. Range: 0.44 to 2.44 hours. Skewness ~0.00 (nearly symmetric).
  • Distribution: Both distance and time are nearly symmetric with very low skewness, confirmed by histograms and skewness statistics in the notebook.

• Categorical Features (from, to, flightType, agency):

  • Cardinality: from (9 unique cities), to (9 unique cities), flightType (3 unique values: 'economic', 'premium', 'firstClass'), agency (3 unique values).
  • Impact on price, distance, and time (Insights from Mean, Mode, and Sum Aggregations via Bivariate Plots in Notebook):
    • flightType:
      • Price: The strongest differentiator.
        • Mean/Median Price: 'firstClass' (e.g., mean ~$1500 from plots) > 'premium' (mean ~$900) > 'economic' (mean ~$800).
        • Mode Price: Similar stratification observed in mode price bar plots.
        • Sum Price (Total Revenue): 'firstClass' generates the most total revenue, followed by 'premium' and 'economic', indicating its high value.
      • Distance/Time: 'firstClass' flights are associated with the highest mean, mode, and sum of distances and times, suggesting they are often longer-haul. 'economic' and 'premium' show similar total distance/time.
    • agency:
      • Price:
        • Mean Price: 'FlyingDrops' shows the highest mean price, followed by 'CloudFy', then 'Rainbow'.
        • Mode Price: 'FlyingDrops' shows higher common prices.
        • Sum Price (Total Revenue): 'CloudFy' and 'Rainbow' are market leaders in total revenue, while 'FlyingDrops' has significantly less, suggesting a niche, higher-margin strategy for FlyingDrops.
      • Distance/Time: 'Rainbow' and 'CloudFy' cover the largest total distance and time. 'FlyingDrops' has the highest mean distance/time.
    • from & to (Routes):
      • Price:
        • Mean/Median Price: Flights involving Salvador (BH) and Florianopolis (SC) (as origin or destination) consistently show higher average prices. Sao Paulo (SP) and Natal (RN) are generally more affordable.
        • Sum Price (Total Revenue): Florianopolis (SC) as a destination and origins like Aracaju (SE), Campo Grande (MS), Brasilia (DF) are among top revenue generators.
      • Distance/Time:
        • Mean/Median Distance/Time: Routes involving Salvador (BH) and Florianopolis (SC) show longer average distances/times.
        • Sum Distance/Time: Florianopolis (SC) (origin and destination) and Aracaju (SE) account for high cumulative flight distances/times.

• Temporal Feature (date and extracted components):

  • Range: Data spans from 2019-09-26 to 2023-07-24.
  • Flight Volume Patterns (from Bar Plots & describe() output):
    • Yearly: Peak flight volume observed in 2020 (55,095 flights in the deduplicated dataset).
    • Monthly: Strong seasonality with the highest flight volumes in Q4 (October, November, December). August shows the lowest volume.
    • Weekly: Thursday is the busiest day with 52,115 flights in the deduplicated dataset. Other days like Monday, Sunday, Friday, Saturday show comparable, lower volumes.
  • Average Price, Time, and Distance Trends (from Line Plots in Notebook):
    • Yearly Trends:
      • Average Price: Shows volatility, with a potential peak around 2020-2021, followed by adjustments.
      • Average Time/Distance: Appear relatively stable across the years with minor fluctuations.
    • Monthly Trends (Seasonality):
      • Average Price: Clear seasonality; peaks during summer (e.g., June-August) and end-of-year (e.g., December). Lower in off-peak months (e.g., February, November).
      • Average Time/Distance: Subtle seasonal pattern, slight increase during peak vacation periods.
    • Weekly Trends (Day-of-Week):
      • Average Price: Fluctuates, commonly higher on Fridays and Sundays. Mid-week often lower. Thursday shows moderate average prices despite high volume.
      • Average Time/Distance: Variations less pronounced; slight tendency for longer average flights on weekends.
  • Action: These strong temporal trends necessitate feature engineering.

• Recommended Feature Engineering (based on notebook actions and EDA insights):

  • Route Definition: Create a route feature by combining from and to.
  • Temporal Features (Implemented & Potential): year, month, day, day_name were extracted. Further consider is_weekend, is_thursday flag, and cyclical encoding for month and day_of_week.
  • Categorical Encoding:
    • flightType, agency: One-hot encoding is suitable.
    • route, from, to: Consider target encoding (with careful cross-validation), frequency encoding, or embedding layers. Cyclic Encoding:
    • month, day
  • Interaction Features: Explore interactions like route_flightType, agency_month, or flightType_day_of_week.

4. 🚨 Outlier Detection & Treatment¶

• Observations:
  • price exhibits right-skewness, with outliers on the higher end. Box plots for categorical features vs. price (e.g., flightType vs. price in the notebook) visually confirm that 'firstClass' and certain agencies/routes contribute more to these higher-priced outliers. time and distance are nearly symmetric and show fewer extreme outliers compared to price.
• Actionable Strategies:
  • Detection: IQR method or Z-scores can be used.
  • Treatment: For price, log transformation is the primary strategy suggested. For other features, if extreme outliers persist and affect sensitive models, consider capping. Tree-based models are generally more robust to outliers.

5. 🕵️ Data Leakage & Bias Risks¶

• Duplicate Data: Mitigated by thorough deduplication after removing non-predictive identifiers.
• Temporal Consistency: Critical: The notebook utilizes a chronological_split function. A time-based split for model training/validation/testing is essential to simulate real-world deployment and prevent temporal leakage, ensuring the model generalizes to future unseen data.

6. 🤖 Modeling Considerations & Strategy¶

• Baseline Model: Tree-based ensembles like Random Forest Regressor or Gradient Boosting Regressor are good starting points due to their robustness to outliers and ability to capture non-linearities.
• Feature Importance & Selection:
  • Primary Drivers: flightType, route (or from/to), agency, and temporal features (month, day_of_week, year) are expected to be highly important based on EDA.
  • Secondary Drivers: distance (or time).
  • Multicollinearity: Given the ~1.0 correlation between distance and time, use only one (e.g., distance) to avoid multicollinearity issues in some models.
• Evaluation Metrics: Standard regression metrics like RMSE (Root Mean Squared Error), MAE (Mean Absolute Error), R² (R-squared), and MAPE (Mean Absolute Percentage Error).
• Iterative Refinement: Plan for hyperparameter tuning, further feature importance analysis, and residual analysis to improve the model.
• Capturing Temporal Dynamics: The model must be structured to effectively learn the strong yearly, monthly, and weekly patterns observed in both flight volume and average price/time/distance.

📋 Summary Table of Key Actions & Strategies¶